When AIs Start Hacking
If you don’t have enough to worry about already, consider a world where AIs are hackers.
Hacking is as old as humanity. We are creative problem solvers. We exploit loopholes, manipulate systems, and strive for more influence, power, and wealth. To date, hacking has exclusively been a human activity. Not for long.
As I lay out in a report I just published, artificial intelligence will eventually find vulnerabilities in all sorts of social, economic, and political systems, and then exploit them at unprecedented speed, scale, and scope. After hacking humanity, AI systems will then hack other AI systems, and humans will be little more than collateral damage.
Okay, maybe this is a bit of hyperbole, but it requires no far-future science fiction technology. I’m not postulating an AI “singularity,” where the AI-learning feedback loop becomes so fast that it outstrips human understanding. I’m not assuming intelligent androids. I’m not assuming evil intent. Most of these hacks don’t even require major research breakthroughs in AI. They’re already happening. As AI gets more sophisticated, though, we often won’t even know it’s happening.
AIs don’t solve problems like humans do. They look at more types of solutions than us. They’ll go down complex paths that we haven’t considered. This can be an issue because of something called the explainability problem. Modern AI systems are essentially black boxes. Data goes in one end, and an answer comes out the other. It can be impossible to understand how the system reached its conclusion, even if you’re a programmer looking at the code.
In 2015, a research group fed an AI system called Deep Patient health and medical data from some 700,000 people, and tested whether it could predict diseases. It could, but Deep Patient provides no explanation for the basis of a diagnosis, and the researchers have no idea how it comes to its conclusions. A doctor either can either trust or ignore the computer, but that trust will remain blind.
While researchers are working on AI that can explain itself, there seems to be a trade-off between capability and explainability. Explanations are a cognitive shorthand used by humans, suited for the way humans make decisions. Forcing an AI to produce explanations might be an additional constraint that could affect the quality of its decisions. For now, AI is becoming more and more opaque and less explainable.
Separately, AIs can engage in something called reward hacking. Because AIs don’t solve problems in the same way people do, they will invariably stumble on solutions we humans might never have anticipated—and some will subvert the intent of the system. That’s because AIs don’t think in terms of the implications, context, norms, and values we humans share and take for granted. This reward hacking involves achieving a goal but in a way the AI’s designers neither wanted nor intended.
Take a soccer simulation where an AI figured out that if it kicked the ball out of bounds, the goalie would have to throw the ball in and leave the goal undefended. Or another simulation, where an AI figured out that instead of running, it could make itself tall enough to cross a distant finish line by falling over it. Or the robot vacuum cleaner that instead of learning to not bump into things, it learned to drive backwards, where there were no sensors telling it it was bumping into things. If there are problems, inconsistencies, or loopholes in the rules, and if those properties lead to an acceptable solution as defined by the rules, then AIs will find these hacks.
We learned about this hacking problem as children with the story of King Midas. When the god Dionysus grants him a wish, Midas asks that everything he touches turns to gold. He ends up starving and miserable when his food, drink, and daughter all turn to gold. It’s a specification problem: Midas programmed the wrong goal into the system.
Genies are very precise about the wording of wishes, and can be maliciously pedantic. We know this, but there’s still no way to outsmart the genie. Whatever you wish for, he will always be able to grant it in a way you wish he hadn’t. He will hack your wish. Goals and desires are always underspecified in human language and thought. We never describe all the options, or include all the applicable caveats, exceptions, and provisos. Any goal we specify will necessarily be incomplete.
While humans most often implicitly understand context and usually act in good faith, we can’t completely specify goals to an AI. And AIs won’t be able to completely understand context.
In 2015, Volkswagen was caught cheating on emissions control tests. This wasn’t AI—human engineers programmed a regular computer to cheat—but it illustrates the problem. They programmed their engine to detect emissions control testing, and to behave differently. Their cheat remained undetected for years.
If I asked you to design a car’s engine control software to maximize performance while still passing emissions control tests, you wouldn’t design the software to cheat without understanding that you were cheating. This simply isn’t true for an AI. It will think “out of the box” simply because it won’t have a conception of the box. It won’t understand that the Volkswagen solution harms others, undermines the intent of the emissions control tests, and is breaking the law. Unless the programmers specify the goal of not behaving differently when being tested, an AI might come up with the same hack. The programmers will be satisfied, the accountants ecstatic. And because of the explainability problem, no one will realize what the AI did. And yes, knowing the Volkswagen story, we can explicitly set the goal to avoid that particular hack. But the lesson of the genie is that there will always be unanticipated hacks.
How realistic is AI hacking in the real world? The feasibility of an AI inventing a new hack depends a lot on the specific system being modeled. For an AI to even start on optimizing a problem, let alone hacking a completely novel solution, all of the rules of the environment must be formalized in a way the computer can understand. Goals—known in AI as objective functions—need to be established. And the AI needs some sort of feedback on how well it’s doing so that it can improve.
Sometimes this is simple. In chess, the rules, objective, and feedback—did you win or lose?—are all precisely specified. And there’s no context to know outside of those things that would muddy the waters. This is why most of the current examples of goal and reward hacking come from simulated environments. These are artificial and constrained, with all of the rules specified to the AI. The inherent ambiguity in most other systems ends up being a near-term security defense against AI hacking.
Where this gets interesting are systems that are well specified and almost entirely digital. Think about systems of governance like the tax code: a series of algorithms, with inputs and outputs. Think about financial systems, which are more or less algorithmically tractable.
We can imagine equipping an AI with all of the world’s laws and regulations, plus all the world’s financial information in real time, plus anything else we think might be relevant; and then giving it the goal of “maximum profit.” My guess is that this isn’t very far off, and that the result will be all sorts of novel hacks.
But advances in AI are discontinuous and counterintuitive. Things that seem easy turn out to be hard, and things that seem hard turn out to be easy. We don’t know until the breakthrough occurs.
When AIs start hacking, everything will change. They won’t be constrained in the same ways, or have the same limits, as people. They’ll change hacking’s speed, scale, and scope, at rates and magnitudes we’re not ready for. AI text generation bots, for example, will be replicated in the millions across social media. They will be able to engage on issues around the clock, sending billions of messages, and overwhelm any actual online discussions among humans. What we will see as boisterous political debate will be bots arguing with other bots. They’ll artificially influence what we think is normal, what we think others think.
The increasing scope of AI systems also makes hacks more dangerous. AIs are already making important decisions about our lives, decisions we used to believe were the exclusive purview of humans: Who gets parole, receives bank loans, gets into college, or gets a job. As AI systems get more capable, society will cede more—and more important—decisions to them. Hacks of these systems will become more damaging.
What if you fed an AI the entire US tax code? Or, in the case of a multinational corporation, the entire world’s tax codes? Will it figure out, without being told, that it’s smart to incorporate in Delaware and register your ship in Panama? How many loopholes will it find that we don’t already know about? Dozens? Thousands? We have no idea.
While we have societal systems that deal with hacks, those were developed when hackers were humans, and reflect human speed, scale, and scope. The IRS cannot deal with dozens—let alone thousands—of newly discovered tax loopholes. An AI that discovers unanticipated but legal hacks of financial systems could upend our markets faster than we could recover.
As I discuss in my report, while hacks can be used by attackers to exploit systems, they can also be used by defenders to patch and secure systems. So in the long run, AI hackers will favor the defense because our software, tax code, financial systems, and so on can be patched before they’re deployed. Of course, the transition period is dangerous because of all the legacy rules that will be hacked. There, our solution has to be resilience.
We need to build resilient governing structures that can quickly and effectively respond to the hacks. It won’t do any good if it takes years to update the tax code, or if a legislative hack becomes so entrenched that it can’t be patched for political reasons. This is a hard problem of modern governance. It also isn’t a substantially different problem than building governing structures that can operate at the speed and complexity of the information age.
What I’ve been describing is the interplay between human and computer systems, and the risks inherent when the computers start doing the part of humans. This, too, is a more general problem than AI hackers. It’s also one that technologists and futurists are writing about. And while it’s easy to let technology lead us into the future, we’re much better off if we as a society decide what technology’s role in our future should be.
This is all something we need to figure out now, before these AIs come online and start hacking our world.
This essay previously appeared on Wired.com
Subscribe to comments on this entry
Léon • April 26, 2021 6:35 AM
No. AI is a misleading term. AI systems are still computers we need to program. Not with a program code this time, but with a model that we need to train.
For a AI to recognise a red car, we need to train the model with images of cars, all sizes and shapes and all colours. And then tell the AI system after each “recognition” whether is was indeed a car and whether it was indeed red or not.
By the way: if one could design an AI for detecting security vulnerabilities, one could also use it in software engineering.